GPT-Live voices: describe one in text, no recording needed
Any API customer can now create a custom GPT-Live voice from a text prompt for free, plus nine new German, French, Hindi, Japanese and Korean presets.
If you build voice apps on the OpenAI API, you can now give your GPT-Live agent its own voice by describing it in plain English. Send the description to /v1/audio/voices (or use the voice Playground), keep the voice ID it returns, and pass that ID when you start a gpt-live-1 session. Creating the voice is free, per OpenAI's changelog.
What changed#
On Oct 16, OpenAI's API changelog said prompt-to-voice is released for all API customers. The same entry added nine natural preset voices for GPT-Live: Nuri, Juni and Flitz (German), Brise and Sillage (French), Nira (Hindi), Harema and Shida (Japanese), and Noeul (Korean). @OpenAIDevs promoted both at 1:08 PM ET.
Before: the only custom voices came from recordings (related: GPT-Live-1.1 in ChatGPT covers the consumer side). You needed a consent recording plus a reference sample, and access is still limited to eligible customers through sales. After: a text description is enough, with no recording and no consent clip.
Why it matters#
Our read: a support line, tutor or booking agent can get a voice that fits the brand in minutes instead of a voice actor session and a sales conversation. Teams shipping in German, French, Hindi, Japanese or Korean also get native presets to start from. Speech still bills as normal GPT-Live session time, which OpenAI lists at $0.05 per minute, billed per second.
How to use it#
- Write a description that covers accent, age, presentation, pitch, timbre, texture, loudness, pace and how distinctive it should be.
- Create the voice with a POST to
/v1/audio/voicesusing"type": "prompt", aname, yourprompt, and"model": "auto". You can add ascript_hintof up to 2,000 characters. - Save the returned
idand keep your original description. - Start a
gpt-live-1session withaudio.output.voiceset to{ "id": "<your voice id>" }, and paste the same description intosession.instructions. - To tune it, change one quality at a time and compare in a fresh session.
Voice description (ours, untested):
A friendly adult voice for a scheduling assistant. Speak with a North American English accent, a mid-range pitch, a clear and lightly textured tone, moderate vocal effort, and a medium, easy pace.
Good to know#
- GPT-Live only. Prompt-made voices do not work in Text to Speech, Realtime or Chat Completions.
- 100 per organization, and that cap includes voices your app creates for its customers. Every create call saves a new voice, so retries use up slots.
- No preview in the response. You hear the voice only by running a session, and Playground listening bills as a session.
- Locked per session. Pick the voice at startup; start a new session to switch.
- Rules. No imitating named people or characters, adult voices only, tell listeners the speech is AI, and OpenAI says voices are not guaranteed to be unique.
- Zero Data Retention. Voice creation through
/v1/audio/voicesis not compatible with it. - Watch next: whether the changelog extends prompt voices to Text to Speech or Realtime.
Sources#
- https://developers.openai.com/api/docs/changelog (Oct 16 entry; Sep 10 GPT-Live GA pricing; Dec 15, 2025 custom voices for eligible customers; checked 2026-10-16 3:57 PM ET)
- https://developers.openai.com/api/docs/guides/custom-voices (checked 3:58 PM ET)
- https://developers.openai.com/api/docs/guides/live-conversations (voice table; checked 3:58 PM ET)
- https://x.com/OpenAIDevs/status/2111142258867335430 (posted 1:08 PM ET / 17:08 UTC Oct 16; checked 3:57 PM ET)
- https://coremesh.online/openai-prompt-to-voice-api/ (secondary, paraphrased only; checked 3:58 PM ET)
Last checked: 2026-10-16